Weighted Memory Tree: +10 Points on GAIA-Text With 33% Fewer Tokens

dair_ai · x · 2026-08-25

A new approach for memory in long-running agents, Weighted Memory Tree, organizes execution into tasks, subtasks, and actions, and assigns every memory a moving retention score — event-based updates raise it, selection-based decay lowers it.

The core trick is "folding": when a subtask finishes, its step-by-step detail collapses into a short summary while the full version stays retrievable; later steps can pull details back. Unlike permanent context trimming, folding delivers the same token savings with a way back.

On GAIA-Text with Qwen3-8B, Gemma 4 E4B, and Llama-3.1-8B, it beats linear memory by 9.97 points on average while using 32.8% fewer prompt tokens. Memory poisoning experiments show retention scoring limits how far unreliable information spreads.

Original post →

More from coding & agent

coding & agent channel →